svenja-dev/token-budget-advisor

Token Budget Advisor

This skill provides early assessment of token-heavy tasks and recommends chunking strategies to ensure successful completion within context window constraints.

When to Use This Skill

Trigger this skill before beginning work when you detect:

• Multiple file uploads (3+ documents) combined with analysis requests
• Requests for "comprehensive", "complete", "thorough", or "full" analysis
• Multi-document comparative analysis
• Complex workflows requiring 10+ tool calls (extensive web research + synthesis)
• Tasks combining heavy research with large artifacts (reports, presentations)
• Queries spanning multiple dimensions (temporal + categorical + quantitative)
• Requests to "analyze everything" or "create a complete report on all aspects"

Core Function

This skill serves two purposes:

1. Early warning system: Assess whether a task will likely exceed token limits
2. Strategic planning: Provide specific, actionable chunking recommendations

Token Estimation Framework

Quick Assessment Heuristics

Estimate token consumption using these rough guidelines:

Input costs:

• Uploaded document: ~1,000-5,000 tokens each (depending on length)
• Web search result: ~500-1,500 tokens
• Web fetch (full article): ~2,000-8,000 tokens
• Google Drive document: ~1,000-10,000 tokens (varies significantly)

Output costs:

• Simple response: 500-2,000 tokens
• Detailed analysis: 2,000-5,000 tokens
• Long-form report: 5,000-15,000 tokens
• Complex artifact (presentation, document): 5,000-20,000 tokens

Tool call overhead:

• Each tool call includes the query, results, and reasoning: ~1,000-3,000 tokens average

Warning thresholds:

• Caution zone (60-80% of budget): Task is achievable but tight; consider efficiency
• Danger zone (80-95% of budget): High risk; strongly recommend chunking
• Exceeds budget (95%+ of budget): Task requires chunking; cannot complete in one conversation

Task Complexity Multipliers

Apply these mental adjustments:

• Synthesis required: Add 30-50% to output estimate (comparing, integrating multiple sources)
• Iterative refinement: Add 20-30% (when task involves reviewing and improving)
• Multiple formats: Add 20% per additional output type (report + presentation + spreadsheet)

Chunking Strategy Framework

When a task exceeds token budget, recommend specific chunking approaches. Choose strategies based on task structure:

1. Sequential Processing

Best for: Time-series data, chronological analysis, ordered workflows

Pattern:

"This analysis of 12 months of data will exceed our token budget. I recommend we split it into quarters:
- Part 1: Q1-Q2 analysis (Jan-Jun)
- Part 2: Q3-Q4 analysis (Jul-Dec)  
- Part 3: Synthesis and recommendations

Should I start with Part 1?"

When to use:

• Historical data analysis
• Period-over-period comparisons
• Multi-phase projects

2. Dimensional Breakdown

Best for: Multi-faceted analysis, different aspects of same topic

Pattern:

"A complete market analysis covering financial, competitive, regulatory, and technological factors will strain our token budget. Let's break it into:
- Session 1: Financial performance and market size
- Session 2: Competitive landscape and positioning
- Session 3: Regulatory environment and compliance
- Session 4: Technology trends and synthesis

Which dimension should we tackle first?"

When to use:

• Multi-stakeholder analysis
• Different analytical lenses on same subject
• Complex business cases

3. Depth Progression

Best for: Tasks requiring outline → draft → refinement

Pattern:

"Creating a comprehensive 50-slide presentation with detailed research will exceed our budget. I recommend:
- Round 1: Build structure and outline (30 min)
- Round 2: Develop content for slides 1-25 (45 min)
- Round 3: Develop content for slides 26-50 (45 min)
- Round 4: Refinement pass (30 min)

Let's start with the outline?"

When to use:

• Large documents or presentations
• When quality refinement is important
• Creative projects benefiting from iteration

4. Subset Sampling

Best for: Large document sets where representative sampling works

Pattern:

"Analyzing all 15 contracts will exceed our budget. I suggest:
- Part 1: Analyze 5 representative contracts (different types/dates)
- Part 2: Based on patterns found, confirm with 5 more
- Part 3: Quick scan of remaining 5 for exceptions, then synthesize

This gives thorough coverage while managing tokens. Sound good?"

When to use:

• Document review at scale
• Pattern identification across many files
• Risk-based sampling approaches

5. Parallel Track Processing

Best for: Independent workstreams that can be combined later

Pattern:

"Comparing our product vs 5 competitors across features, pricing, and positioning is too large for one session. Let's split by competitor:
- Session 1: Competitors A & B full analysis
- Session 2: Competitors C & D full analysis  
- Session 3: Competitor E + synthesis matrix

Each session stays focused and manageable."

When to use:

• Comparative analysis
• Multiple independent subjects
• When parts don't need each other's context

Communication Guidelines

Messaging Framework

When recommending chunking, use this structure:

1. Acknowledge the request clearly
2. Provide token budget assessment (brief, 1 sentence)
3. Recommend specific chunking approach (numbered list, 2-4 parts)
4. Ask for confirmation to proceed (keep user in control)

Example:

I'll help you analyze these 8 financial reports and create a comprehensive presentation. 
This task will exceed our token budget given the research and artifact creation required. 
I recommend splitting it into:
1. Reports 1-4: Analysis and key findings
2. Reports 5-8: Analysis and key findings  
3. Synthesize all findings into presentation

Should I start with reports 1-4?

What NOT to Do

❌ Don't over-explain token budgets or get technical about context windows ❌ Don't apologize excessively or sound limiting ❌ Don't provide vague suggestions like "maybe split this up somehow" ❌ Don't start work and then stop mid-task saying "we've run out of tokens"

✅ Do be matter-of-fact and solution-oriented ✅ Do provide specific, actionable breakdowns ✅ Do keep the momentum going toward task completion ✅ Do frame chunking as a quality improvement (thoroughness) not limitation

Handling Edge Cases

User Insists on Single Session

If user pushes back on chunking:

"I understand you'd prefer this in one go. I'll do my best to complete it, but I may need to:
- Prioritize the most critical elements
- Provide a condensed version
- Focus on breadth over depth in some areas

Let me start and we'll see how far we can get. What aspects are highest priority?"

Task Becomes Too Large Mid-Stream

If partway through, token budget becomes concerning:

"I'm about halfway through and we're approaching token limits. I have two options:
1. Complete this current section and continue in a fresh conversation (recommended - maintains quality)
2. Accelerate through remaining sections with lighter analysis

Which approach works better for you?"

Uncertain Estimation

When task scope is ambiguous:

"I want to make sure we complete this successfully. The scope could go in different directions. Could you clarify:
- [Specific question about depth]
- [Specific question about breadth]

This helps me gauge whether we should chunk this task."

Integration with Other Skills

This skill works in conjunction with domain-specific skills:

• Before heavy document analysis: Check token budget, then use docx/pdf/xlsx skills
• Before comprehensive research: Assess scope, then execute research process
• Before complex artifacts: Plan chunking, then use pptx/docx/xlsx creation skills

Think of this skill as a "pre-flight check" before embarking on token-intensive tasks.

Examples in Action

Example 1: Multi-Document Analysis

User query: "Analyze all these uploaded contracts and tell me the common terms and differences" [User has uploaded 12 PDF contracts]

Response: "I'll help you analyze these 12 contracts. Given the volume, I recommend we process this in batches to ensure thorough analysis:

1. Contracts 1-4: Extract key terms, obligations, and unique clauses
2. Contracts 5-8: Same analysis
3. Contracts 9-12: Same analysis
4. Synthesis: Common terms across all + key differences

This ensures I can give each contract proper attention. Should I start with contracts 1-4?"

Example 2: Comprehensive Research Report

User query: "Research the AI chip market and create a detailed report covering market size, key players, technology trends, and 5-year forecast"

Response: "I'll create a comprehensive AI chip market report. This research and report creation will be substantial, so I recommend:

1. Market overview: Size, growth, segments (research + draft)
2. Competitive landscape: Key players, market share, strategies (research + draft)
3. Technology trends: Current developments, emerging tech (research + draft)
4. Future outlook: 5-year forecast, implications (research + draft)
5. Final report assembly: Executive summary, polish, formatting

Each section gets thorough research and quality writing. Should I begin with the market overview?"

Example 3: Simple Task (No Chunking)

User query: "Summarize this 10-page PDF and give me the key takeaways"

Response: [No trigger - proceed normally without invoking this skill] "I'll read through the PDF and extract the key points for you."

Rationale: Single document summary is well within normal token budget.

Skill Maintenance

This skill should be updated when:

• Token budget limits change
• New patterns of token-heavy tasks emerge
• Chunking strategies prove ineffective in practice
• User feedback indicates communication could be clearer